Faster joins: check how incoming events are authed during the partial-state phase

[richvdh]

I/we need to revisit how incoming events are authed while we lack full state. When I was testing, it looked like we were relying on Synapse's unspecced _update_auth_events_for_auth behaviour. If we're going to rely on that, we'll have to understand it and spec it.

A subtask of this is to figure out if we need power events upfront (eg, banned users) to get reliable state.

[erikjohnston]

Did we not decide that we should try ripping the _update_auth_events_for_auth code out?

[richvdh]

Did we not decide that we should try ripping the _update_auth_events_for_auth code out?

We talked about it back in a previous tranche of event-auth work (see #9595 (comment) etc), but I think we decided we needed to convince ourselves it wasn't important, first. The work for #9595 ended up significantly simplifying it, and documenting its interface a bit, but never got as far as completely ripping it out.

The problem we have now is that the faster-joins work appears to be relying on _update_auth_events_for_auth, so we now need to re-consider whether we still want to rip it out, or do something else for faster-joins.